Output is correct in EmuDebug mode but wrong in Debug mode

Accelerated Computing CUDA CUDA Programming and Performance
#1

Hi,

I am facing a problem with my code

CPU CODE

char *** symbol;
int *** tree;
// Allocate memory to tree and symbol;

dp<<<1,3>>>(tree,symbol,threshold);

GPU CODE

global void dp(int ***tree, char **symbol1,int threshold)
{
int ind_tree=blockDim.xblockIdx.x+threadIdx.x;
for(int i=0;i<3;i++)
{
for(int j=0;j<pow(4.0,i);j++)
…
…
// value of symbol1 updated in order to update symbol on CPU
…
…
}

The code works fine when I run in EmuDebug mode however when ran otherwise the code messes up somewhere in the kernel.
The values that I get back in ***symbol are same as that before the kernel call. The values of tree however are correctly reflected after the kernel call is over.

#2

Your code fragment explains very little.

It is very likely you are passing host memory, or pointers to host memory, or pointers to pointers to host memory, to your kernel.

#3

Yes I am passing a pointer to host memory. The data structure is arranged as a tree. The dp is a call to do a dominant pass over one tree. I am attaching the .cu file if that can help. The troublesome kernel call is in line 207 and the kernel code starts from line 124. The weird thing is that a very similar data structure in the code (int *** tree) is getting correctly updated.

code.txt (6.97 KB)

#4

A kernel can only be called with a pointer to global memory.

#5

:oops: am so very sorry… I meant global memory only… Very sorry again.

Also I am getting an exception when the program terminates that says

First-chance exception at 0x76b742eb in ezw_cuda16.exe: Microsoft C++ exception: cudaError_enum at memory location 0x0012cdd8…

I am getting this exception only some of the times and not for every run of the code.

#6

My problem is quite simple as I now see it. Can someone please explain as to why pointer p is not getting updated in debug mode? *p is getting updated in emudebug mode but is not happening in debug mode. Thank you.

#include <stdio.h>
#include <stdlib.h>
#include <cuda_runtime.h>
#include <conio.h>
char *p;
global void updatep(char p1)
{
p1=‘e’;
}
int main(int argc, char argv)
{
p=(char *)malloc(sizeof(char));
*p=‘a’;
printf("%c",*p);
updatep<<<1,1>>>§;
printf("\n%c",*p);
getch();
return 0;
}

#7

How is the GPU going to update a value held on host memory? This works in emulation because then the GPU and host memory are not distinct.

#8 
#include <stdio.h>

#include <stdlib.h>

#include <cuda_runtime.h> 

#include <conio.h>

char *p;

__global__ void updatep(char *p1)

{

*p1='e';

}

int main(int argc, char** argv)

{

	p=(char *)malloc(sizeof(char));

	*p='a';

	printf("%c",*p);

// Here you need to allocate global memory on gpu, copy p to that allocated memory.

	updatep<<<1,1>>>(p);

// Here you need to copy global memory from gpu back to the host.

	printf("\n%c",*p);

	getch();

	return 0;

}
#9

I want to thank you for your patience for helping me solve my mistakes. I had to modify my code to make everything work in one dimension. I am happy that now the code is running fine. Thanks for your time.